Pith. sign in

REVIEW 3 major objections 5 minor 37 references

3D Trajectory Reconstruction of Moving Points Based on Asynchronous Cameras

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that a moving point's 3D trajectory can be reconstructed from asynchronous cameras by jointly optimizing target motion parameters, camera time offsets, frame rates, and camera rotations in a single bundle-adjustment…

desk verdict A solid, incremental bundle adjustment formulation for asynchronous trajectory reconstruction, but the headline field result is really a closed-loop reprojection experiment, so read the 112.95 m claim with that caveat. read the letter →

arxiv 2506.00541 v2 pith:6ZTDRIZN submitted 2025-05-31 cs.CV

Huayu Huang , Banglei Guan , Yang Shang , Qifeng Yu This is my paper
classification cs.CV
keywords 3Dtrajectoryreconstructionasynchronouscamerastemporalpolynomialsbundleadjustmentintersectioncamerasynchronizationtimeoffsetestimationphotomechanics
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tackles a practical measurement problem: multiple cameras on separate platforms rarely capture the same moving point at the same instant, so ordinary triangulation fails. It claims that if the target's short-term motion is described by a low-order temporal polynomial, the camera time offsets (and even frame rates and rotations) can be estimated in the same optimization that reconstructs the trajectory. The payoff is a single bundle-adjustment step instead of iterating between synchronization and reconstruction, with higher accuracy in simulations and a real-world localization error of 112.95 m at an observation range of 15–20 km. A reader should care because it removes the need for hardware synchronization in UAV-based photomechanical measurement.

What carries the argument

The load-bearing object is the temporal-polynomial trajectory model coupled with the perpendicular-residual sight-ray constraint. Each camera's frame time is written as $t_c(f)=f/\alpha_c+\beta_c$, and the target position at that time is expanded as polynomials of degree $K$ in $t$ (Eq. 7). The residual projects the difference between the model position and the camera center onto the plane perpendicular to the sight ray (Eq. 5); stacking these residuals across cameras, frames, and targets yields a bundle-adjustment objective (Eq. 14) whose unknowns are motion coefficients, time offsets $\beta_c$, frame rates $\alpha_c$, and quaternion-parameterized rotations. The trajectory intersection therefore replaces triangulation, and the joint bundle adjustment replaces iterative synchronization.

What would settle it

Run the method on a target that performs a sharp turn or rapid acceleration inside the observation window, with GNSS ground truth; if the localization error becomes much larger than for smooth motion under identical noise and the rotation refinement does not compensate, the polynomial motion model is the limiting factor.

Watch

Extended reading notes

Core claim

The central claim is that the traditional triangulation constraint, which requires synchronized cameras, can be replaced by a trajectory-intersection residual evaluated at per-camera times. If the target trajectory is written as temporal polynomials in a global time variable, then the unknown polynomial coefficients, the cameras' time offsets, the frame rates, and (when multiple targets are present) the camera rotations can all be optimized together in a single nonlinear least-squares problem. The paper demonstrates in simulation that this joint optimization is more accurate and roughly twice as fast as the iterative method of Li et al., and that optimizing rotations from moving-point constraints substantially reduces error when rotation measurements are poor. In the real-world test, the method reconstructs a highway vehicle observed from 15–20 km with a 3D localization error of 112.95 m.

Load-bearing premise

The target's motion within the observation window must be well approximated by a low-order temporal polynomial (uniform or uniformly accelerated motion), and the paper's experiments only test smooth, near straight-line motion.

Editorial extensions

If this is right

  • No hardware synchronization is needed: any set of cameras with known intrinsics and rough pose can reconstruct a moving point, as long as the trajectory is smooth enough to be a low-order polynomial.
  • Inaccurate frame rates no longer degrade reconstruction, because the frame rate is an optimization variable rather than a fixed input.
  • When several moving targets are visible, the extra geometric constraints improve the camera rotation estimates, which is the dominant error source on platforms with low-grade IMUs.
  • The method inherits the occlusion tolerance of trajectory intersection: targets do not need to be visible in every camera at every frame, and enough constraints are counted by the solvability condition.
  • The polynomial coefficients carry physical meaning (position, velocity, acceleration), so the same optimization yields kinematic characteristics of the target.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural next stress test is to apply the same joint optimization to a heterogeneous camera network (different frame rates, rolling shutters, or mixed EO/IR sensors), where the temporal model would need adjustment but the core residual structure should transfer.
  • The polynomial order K is chosen by hand; a data-driven selection of K or a switch to basis splines would likely extend the method to maneuvering targets, which the current experiments do not exercise.
  • The coplanar-sight-ray degeneracy noted for unknown frame rates and offsets could be turned into a design rule for UAV observation geometry: plan flight paths that keep sight rays non-coplanar.
  • Because the residual is perpendicular to the sight ray, the formulation should generalize to non-point features such as line or area observations, with the temporal polynomial describing the object's motion rather than a centroid.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper addresses 3D trajectory reconstruction of moving points from asynchronous cameras. It extends the trajectory intersection method to multi-camera asynchronous settings, represents target motion with temporal polynomials, and proposes bundle adjustment frameworks that jointly optimize target motion parameters, camera time information (frame rate and offset), and—when multiple targets are observed—per-frame camera rotations. The method is evaluated in extensive simulations (1,000 trials per configuration) and in a real-world experiment in which a virtual second camera is constructed by reprojection with added noise, yielding a claimed localization error of 112.95 m at 15–20 km range.

Significance. If the claims are fully supported, the work is a useful contribution: it solves two coupled sub-problems (trajectory reconstruction and camera synchronization) in one optimization, outperforms the iterative baseline Li et al. in accuracy and speed, and improves accuracy when frame rate or rotation measurements are inaccurate. The simulation study is extensive and carefully repeated. However, the headline real-world accuracy is not an independent field measurement because the second view is generated by reprojection from known ground truth, and the motion model is only tested on smooth, near-constant-velocity trajectories. These limitations materially narrow the strength of the empirical claims.

major comments (3)
  1. [3.2, Table 3] The claimed real-world validation is not an independent measurement of an asynchronous camera system. As stated in Section 3.2, the virtual camera is constructed by reprojection and its target image points are obtained by reprojection from the known GNSS trajectory and platform poses, with noise added. The same known trajectory and poses are then used as ground truth for the error. Therefore the 112.95 m error in Table 3 and the corresponding statements in the abstract and conclusion do not establish the method's accuracy on real asynchronous cameras. Independent effects such as per-view tracking errors, independent IMU rotation errors, platform vibration, and unknown time offsets that cannot be manufactured by reprojection are absent. I recommend either obtaining genuine two-camera data or clearly reclassifying this experiment as a simulation and adjusting the claims accordingly.
  2. [2.2, Eq. (7); Section 3 experiments] The central motion model assumes that the target trajectory is a low-order temporal polynomial (K=1 or 2) over the observation window. All simulation targets and the real target move in near-uniform straight-line or uniformly accelerated motion, so the experiments do not stress-test what happens when the target maneuvers strongly within the window. Under model misspecification, the joint optimization of time offsets, rotations, and motion parameters may absorb the bias in a way that degrades both trajectory and synchronization accuracy. The paper should either add experiments with maneuvering targets (e.g., piecewise-constant acceleration or sinusoidal deviations) or explicitly scope the method and claims to smooth, short-window motions for which the polynomial assumption is reasonable.
  3. [2.4, solvability discussion] Algorithm 2, which optimizes camera rotations, is unsolvable when there is only one target, as the paper states. In the real-world experiment this limitation is circumvented by adding three virtual targets around the actual target, but those virtual targets are synthetic and their image points are generated by reprojection. Consequently, the improvement of Algorithm 2 over Algorithm 1 shown in Table 3 is not demonstrated on real data; it reflects constraints supplied by a simulated multi-target scenario. The paper should clarify that the real-world demonstration of rotation optimization is semi-synthetic and should ideally validate Algorithm 2 on real multi-target observations or a real second camera.
minor comments (5)
  1. [Eq. (1)] The projection equation is typeset incorrectly: the matrix containing R_c(f), T_c(f), and the row "1 1" is not the standard [R|t] augmented matrix, and the notation is confusing. Please rewrite it in the conventional form [x;1] ~ K_c [R_c(f) | t_c(f)] [X;1].
  2. [2.2, after Eq. (8)] The text says the number of independent equations is 2NC, but the variable N is not defined there; it should be 2FC where F is the number of frames. Also, the sentence describing the reconstructed trajectory as "linear polynomials of time" should say "temporal polynomials of order K," since K is general.
  3. [2.4, parameter count] The total number of optimized parameters is stated as sum 3(K_n+1) + 3 sum F_c + C, but the time-information parameters number 2C. The final term should be 2C to be consistent with the solvability inequality that follows.
  4. [3.1, Figure 7] The noise levels for the high- and low-noise cases in the offset-estimation experiment are not specified in the text; the reader cannot tell what standard deviations were used. Please report these values.
  5. [3.2] The description of the real-world scene says the virtual platform trajectory shares the same shape as the flight platform, but the distinction between the real platform's pose (provided by GNSS/IMU) and the virtual platform's presumed pose is not explicit. Clarifying the coordinate frame and how the 20 km baseline is realized would improve reproducibility.

Circularity Check

0 steps flagged · score 2.0 of 10

Core derivation is self-contained; minor self-citations and a closed-loop virtual-camera validation limit the empirical claim but do not make the estimation circular.

full rationale

The derivation chain in Sections 2.2–2.4 is built from the projection residual (Eqs. (5)–(6)), the camera time model (Eq. (2)), and an explicit polynomial motion model (Eqs. (7), (10), (12)). The unknowns—motion parameters a_k, b_k, c_k, frame rates α_c, offsets β_c, and rotations q_c(f)—are jointly minimized in the objective functions (Eqs. (9), (11), (14)); none of these outputs is defined as an input, and no fitted quantity is later relabeled as a prediction. The temporal-polynomial ansatz is imported from the same group's earlier trajectory-intersection work [23], but the paper restates the model and gives a physical rationale, and it invokes no uniqueness theorem to forbid alternatives; this is a minor self-citation that is not load-bearing. The strongest empirical result (112.95 m at 15–20 km, Table 3) must be read with the disclosed limitation in Section 3.2: the 'real-world' experiment uses a monocular camera plus a virtual camera whose image points are obtained by reprojection from the GNSS ground-truth trajectory and platform pose, with noise added. That makes the validation partly closed-loop and weakens the evidential force of the headline number, but it is a validation limitation rather than a circular derivation: the reported error is not forced to a target value by the construction, and the central estimation problem remains one of fitting observations through a geometric residual. No derivation step reduces to its own inputs, so the overall circularity score is low.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The method's solvability depends on a smooth-motion prior (temporal polynomials), calibrated intrinsics, and known camera positions. The real-world validation relies on a virtual second camera, which is a modeling construct rather than an independent sensor. These are enumerated above; the method itself also introduces many per-frame rotation parameters that are fitted, though they are part of the intended optimization.

free parameters (4)
  • Temporal polynomial order K = 1 (linear) or 2 (accelerated) in experiments
    Chosen by hand based on target dynamics; determines the number of motion parameters 3(K+1) and the smoothness of the reconstructed trajectory.
  • Camera time offset beta_c = 0.1 to 1.0 frames in simulation; 40 frames in the real-world setup
    The central unknown the method estimates; the paper admits convergence to a joint optimum of offset and motion parameters rather than the global time offset, so estimates carry bias.
  • Per-frame camera rotations q_c(f) = Optimized in every frame for Algorithm 2
    The framework adds 3 degrees of freedom per frame per camera, which can absorb projection errors and may overfit when only a few targets are observed.
  • Frame rate alpha_c = 10 Hz simulated, 25 Hz real; tested with wrong inputs of 9 Hz and 20 Hz
    Treated as an optimization variable when inaccurate; the reported robustness depends on optimizing this value.
assumptions (6)
  • domain assumption Target trajectory can be represented by low-order temporal polynomials over the observation window
    Introduced in Sect. 2.2, Eq. (7); order K chosen by hand (1 or 2). The whole underdetermined problem becomes solvable only under this smoothness prior.
  • domain assumption Camera intrinsic parameters are calibrated in advance and held fixed
    Stated after Eq. (1); no online estimation of intrinsics.
  • domain assumption Camera translations (optical centers) are known from GNSS or SfM
    Used in Eq. (3) and throughout; only rotations are optimized, not positions.
  • standard math Point-to-line residual in Eq. (5) is the correct geometric objective
    Standard geometric error for sight-ray intersection; reasonable but a modeling choice.
  • domain assumption Levenberg-Marquardt with beta initialized to 0 converges to a useful minimum
    Sect. 2.3; paper notes it may reach a joint optimum rather than global time offset, so initialization and local minima are a risk.
  • ad hoc to paper A virtual camera created by reprojection with known geometry faithfully represents a real asynchronous second camera
    Sect. 3.2, Figure 12; the 'real-world' binocular setup uses a synthetic second view, which shapes the headline error.

how reviews work

0 comments
Cite this review

Pith. "Pith review of 3D Trajectory Reconstruction of Moving Points Based on Asynchronous Cameras." pith.science (2026). https://pith.science/paper/6ZTDRIZN

@misc{pith2026250600541,
  author       = {Pith},
  title        = {Pith review of: 3D Trajectory Reconstruction of Moving Points Based on Asynchronous Cameras},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6ZTDRIZN}},
  note         = {Machine review of arXiv:2506.00541}
}
read the original abstract

Photomechanics is a crucial branch of solid mechanics. The localization of point targets constitutes a fundamental problem in optical experimental mechanics, with extensive applications in various missions of UAVs. Localizing moving targets is crucial for analyzing their motion characteristics and dynamic properties. Reconstructing the trajectories of points from asynchronous cameras is a significant challenge. It encompasses two coupled sub-problems: trajectory reconstruction and camera synchronization. Present methods typically address only one of these sub-problems individually. This paper proposes a 3D trajectory reconstruction method for point targets based on asynchronous cameras, simultaneously solving both sub-problems. Firstly, we extend the trajectory intersection method to asynchronous cameras to resolve the limitation of traditional triangulation that requires camera synchronization. Secondly, we develop models for camera temporal information and target motion, based on imaging mechanisms and target dynamics characteristics. The parameters are optimized simultaneously to achieve trajectory reconstruction without accurate time parameters. Thirdly, we optimize the camera rotations alongside the camera time information and target motion parameters, using tighter and more continuous constraints on moving points. The reconstruction accuracy is significantly improved, especially when the camera rotations are inaccurate. Finally, the simulated and real-world experimental results demonstrate the feasibility and accuracy of the proposed method. The real-world results indicate that the proposed algorithm achieved a localization error of 112.95 m at an observation range of 15 ~ 20 km.

Figures

Figures reproduced from arXiv: 2506.00541 by the authors.

Figure 3
Figure 3. (i) The order of the temporal polynomials represent￾ing the camera motion is equal to or lower than that of the target motion, as shown in Fig. 4a. (ii) All sight-rays intersect at the same point (all sight￾rays being parallel is a special case of this situation, where they intersect at the infinite point), as shown in Fig. 4b and 4c [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 37 canonical work pages

  1. [1]

    Y. Fu, Y. Shang, W. Hu, B. Li, and Q. Yu, Non-contact optical dynamic measurements at different ranges: a review. Acta Mech. Sin. https://doi.org/10.1007/s10409-021-01102-1 37: 537–553 (2021)

    work page doi:10.1007/s10409-021-01102-1 2021

  2. [2]

    Z. Liu, B. Guan, Y. Shang, Q. Yu and L. Kneip, Line-Based 6-DoF Object Pose Estimation and Tracking With an Event Camera, IEEE Trans. Image Process. 33: 4765-4780 (2024)

    work page 2024

  3. [3]

    L. Gao, Y. Liu, B. Jiang, Z. Ge, H. Li, H. Guo, T. Suo, and Q. Yu, A novel videogrammetry-based full-field dynamic deformation monitoring method for variable-sweep wings. Acta Mech. Sin. https://doi.org/10.1007/s10409-024-23639-x 40: 423639 (2024)

    work page doi:10.1007/s10409-024-23639-x 2024

  4. [4]

    Z. Liu, B. Guan, Y. Shang, Y. Bian, P. Sun and Q. Yu, Stereo Event-Based, 6-DOF Pose Tracking for Uncooperative Spacecraft, IEEE Trans. Geosci. Remote Sens. 63: 1-13 (2025)

    work page 2025

  5. [5]

    C. Zhao, H. Yu, M. Jiang, and X. Tian, PSMNet: A Neural Network-Driven Approach for Pixel Similarity Measurement in Distributed Scatterer Interferometry, IEEE Trans. Geosci. Remote Sens. (2025)

    work page 2025

  6. [6]

    Tian and S

    B. Tian and S. Bhattacharya, Modelling and control of a spatial dynamic cable. Acta Mech. Sin. https://doi.org/10.1007/s10409-019-00844-3 35: 866–878 (2019)

    work page doi:10.1007/s10409-019-00844-3 2019

  7. [7]

    Huang, C

    H. Huang, C. Chen, B. Guan, Z. Tan, Y. Shang, Z. Li, and Q. Yu, Ridge estimation-based vision and laser ranging fusion localization method for UAVs, Appl. Opt. 2025, 64(6): 1352-1361

    work page 2025

  8. [8]

    Liang, S

    H. Liang, S. C. Lee, and S. Seo, UAV-based low altitude remote sensing for concrete bridge multi-category damage automatic detection system, Drones, 2023, 7(6): 386

    work page 2023

Show all 37 references
  1. [9]

    S. Feng, M. Gao, X. Jin, T. Zhao, and F. Yang, Fine-grained damage detection of cement concrete pavement based on UAV remote sensing image segmentation and stitching, Measurement, 2024, 226: 113844

    2024

  2. [10]

    H. Hu, B. Wei, S. Mei, J. Liang, and Y. Zhang, A two-step calibration method for vision measurement with large field of view, IEEE Trans. Instrum. Meas. 71: 1–10 (2022)

    2022

  3. [11]

    Magera, T

    F. Magera, T. Hoyoux, O. Barnich, and M. V. Droogenbroeck, A universal protocol to benchmark camera calibration for sports, IEEE conference on computer vision and pattern recognition, Seattle, 2024

    2024

  4. [12]

    Liang, B

    S. Liang, B. Guan, Z. Yu, P. Sun, and Y. Shang, Camera calibration using a collimator system, European Conference on Computer Vision, Milan, 2024

    2024

  5. [13]

    Hartley, and A

    R. Hartley, and A. Zisserman, Multiple view geometry in computer vision (2nd ed.), Cambridge university press, 2004

    2004

  6. [14]

    H. C. Longuet-Higgins, A computer algorithm for reconstructing a scene from two projections, Nature 293 (5828) (1981) 133–135

    1981

  7. [15]

    Y. Bian, H. Zhao, Z. Liu, Z. Su, B. Guan, D. Zhang, and Q. Yu, Stereo vision-based health monitoring method for wind turbine towers, Measurement 226 (2024) 114148

    2024

  8. [16]

    S. H. Lee, and J. Civera, Triangulation: why optimize?, arXiv preprint arXiv:1907.11917 (2019)

    2019 arXiv

  9. [17]

    M. Vo, S. G. Narasimhan, and Y. Sheikh, Spatiotemporal bundle adjustment for dynamic 3D reconstruction, IEEE conference on computer vision and pattern recognition, Las Vegas, 2016

    2016

  10. [18]

    Avidan, and A

    S. Avidan, and A. Shashua, Trajectory triangulation: 3D reconstruction of moving points from a monocular image sequence, IEEE Trans. Pattern Anal. Mach. Intell. 22 (4) (2000) 348–357

    2000

  11. [19]

    J. Y. Kaminski, and M. Teicher, A general framework for trajectory triangulation, J. Math. Imaging Vis. 21(1): 27–41 (2004)

    2004

  12. [20]

    Akhter, Y

    I. Akhter, Y. Sheikh, S. Khan, and T. Kanade, Trajectory space: A dual representation for nonrigid structure from motion, IEEE Trans. Pattern Anal. Mach. Intell. 33 (7):1442–1456 (2011)

    2011

  13. [21]

    Y. Zhu, M. Cox, and S. Lucey, 3D motion reconstruction for real-world camera motion, IEEE Conference on Computer Vision and Pattern Recognition, Colorado, 2011

    2011

  14. [22]

    H. S. Park, T. Shiratori, I. Matthews, and Y. Sheikh, 3D trajectory reconstruction under perspective projection, Int. J. Comput. Vis. 115 (2):115–135 (2015)

    2015

  15. [23]

    Q. Yu, Y. Shang, J. Zhou, X. Zhang, and L. Li, Monocular trajectory intersection method for 3D motion measurement of a point target, Sci. China Ser. E-Technol. Sci. 52, 3454–3463 (2009)

    2009

  16. [24]

    X. Li, P. Guo, H. Zhang, and X. Zhang, LS-SVM for a point target motion measurement, International Conference on Signal Processing, Beijing, 2014

    2014

  17. [25]

    S. Chen, H. Liu, X. Liu, and Q. Yu, Non-cooperative maritime target position and velocity measuring method based on monocular trajectory intersection for video satellite, Proc. IMechE. Part G: J. Aerospace Eng. 233(1): 44–56 (2019)

    2019

  18. [26]

    Caspi, D

    Y. Caspi, D. Simakov, and M. Irani, Feature-based sequence-to-sequence matching, Int. J. Comput. Vis. 68: 53-64 (2006)

    2006

  19. [27]

    Elhayek, C

    A. Elhayek, C. Stoll, K. I. Kim, H. P. Seidel, and C. Theobalt, Feature-based multi-video synchronization with subframe accuracy, Joint DAGM (German Association for Pattern Recognition) and OAGM Symposium, Berlin, 2012

    2012

  20. [28]

    F. L. Padua, R. L. Carceroni, G. A. Santos, and K. N. Kutulakos, Linear sequence-to-sequence alignment, IEEE Trans. Pattern Anal. Mach. Intell. 32(2): 304-320 (2008)

    2008

  21. [29]

    P. A. Tresadern, and I. D. Reid, Video synchronization from human motion using rank constraint, Comput. Vis. Image Understanding, 113(8): 891-906 (2009)

    2009

  22. [30]

    C. Albl, Z. Kukelova, A. Fitzgibbon, J. Heller, M. Smid, T. Pajdla, On the two-view geometry of unsynchronized cameras, IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, 2017

    2017

  23. [31]

    X. Zhou, Y. Dai, H. Qin, S. Qiu, X. Liu, Y. Dai, J. Li, and T. Yang, Subframe-Level Synchronization in Multi-Camera System Using Time-Calibrated Video, Sensors, 24(21): 6975 (2024)

    2024

  24. [32]

    X. Li, J. Liu, X. Yang, and Q. Yu, Multi-camera trajectory triangulation for point target motion measurement in different conditions of time information, J. Natl. Univ. Def. Technol. 37(3): 110-114 (2015)

    2015

  25. [33]

    M. Vo, Y. Sheikh, and S. G. Narasimhan, Spatiotemporal bundle adjustment for dynamic 3d human reconstruction in the wild, IEEE Trans. Pattern Anal. Mach. Intell. 44(2): 1066-1080 (2020)

    2020

  26. [34]

    H. S. Park, T. Shiratori, I. Matthews, and Y. Sheikh, 3D reconstruction of a moving point from a series of 2d projections, European Conference on Computer Vision, Hersonissos, 2010

    2010

  27. [35]

    Zhang, Y

    B. Zhang, Y. Yang, and S. Liao, Ultra-chaos of square thin plate in low Earth orbit. Acta Mech. Sin. https://doi.org/10.1007/s10409-023-23428-x 40, 523428 (2024)

    2024 doi

  28. [36]

    F. Gao, J. Li, F. Dong, J. Yi, and G. Sun, Dynamic modeling and control strategy for tethered satellite systems in orbital debris management. Acta Mech. Sin. https://doi.org/10.1007/s10409-024-24262-x 41, 524262 (2025)

    2025 doi

  29. [37]

    J. F. Henriques, R. Caseiro, P. Martins, and J. Batista, High-speed tracking with kernelized correlation filters, IEEE Trans. Pattern Anal. Mach. Intell. 37(3): 583–596 (2015)

    2015

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.